Unsupervised data classification using pairwise Markov chains with automatic copulas selection

The Pairwise Markov Chain (PMC) model assumes the couple of observations and states processes to be a Markov chain. To extend the modeling capability of class-conditional densities involved in the PMC model, copulas are introduced and the influence of their shape on classification error rates is studied. In particular, systematic experiments show that the use of wrong copulas can degrade significantly classification performances. Then an algorithm is presented to identify automatically the right copulas from a finite set of admissible copulas, by extending the general “Iterative Conditional Estimation” (ICE) parameters estimation method to the context considered. The unsupervised segmentation of a radar image illustrates the nice behavior of the algorithm.

► Introduction of copulas in the Pairwise Markov Chain model. ► Unsupervised parameters estimation of parameters by means of ICE. ► Automatic copula selection within a finite set of candidates.